Researchers question AI capability acceleration: shift to RL may be the confounder
DavidSKrueger · x · 2026-11-09
David Krueger (Cambridge) signal-boosts a critique of claimed AI capability acceleration.
- Original evidence: accelerating internal benchmark trends and researcher productivity without acceleration in total compute
- @GeorgeE33179 points out the most obvious confounding variable: training compute shifting towards RL
- That alone, he argues, explains all three pieces of evidence without invoking a qualitative capability jump
Krueger finds the critique substantive enough to merit a formal response.
More from Models
- Thinky Inkling 2 Small: New Open US Model Trails DeepSeek Flash While Costing More — intellectronica · 2026-11-09
- Gemini 4 Carbon Ranks #1 on Artificial Analysis but No One Can Actually Use It — Yuchenj_UW · 2026-11-09
- Users report inconsistent AI bot allowance usage and weekend performance dips — mattyp · 2026-11-09
- Uncensored 9B distill 'DeepSeek-V4-Pro-Qwen3.5-9B' claims to run locally on a 5GB laptop — udmrzn · 2026-11-09
- Garry Tan's open-source eval: 7 coding harnesses tie on capability but differ 2x on cost — garrytan · 2026-11-09
- "Opus 5.5 is an artist trapped in an LLM": community raves over creative coding output — alejandroll10 · 2026-11-09